Details

Autor(en) / Beteiligte

• Vieira, Miguel Ângelo Conde

Titel

Development of a Corpus for User­Based Scientific Question Answering

Ort / Verlag

ProQuest Dissertations & Theses

Erscheinungsjahr

2021

Link zum Volltext

• ProQuest

Quelle

ProQuest Dissertations & Theses A&I

Beschreibungen/Notizen

• In recent years Question & Answering (QA) tasks became particularly relevant in the research field of natural language understanding. However, the lack of good quality datasets has been an important limiting factor in the quest for better models. Particularly in the biomedical domain, the scarcity of gold standard labelled datasets has been a recognized obstacle given its idiosyncrasies and complexities often require the participation of skilled domain¬specific experts in producing such datasets.To address this issue, a method for automatically gather Question¬Answer pairs from online QA biomedical forums has been suggested yielding a corpus named BiQA. The authors describe several strategies to validate this new dataset but a human manual verification has not been conducted.With this in mind, this dissertation was set out with the objectives of performing a manual verification of a sample of 1200 questions of BiQA and also to expanding these questions, by adding features, into a new corpus of text ¬ BiQA2 ¬ with the goal of contributing with a new corpus for biomedical QA research. Regarding the manual verification of BiQA, a methodology for its characterization was laid out and allowed the identification of an array of potential problems related to the nature of its questions and answers aptness for which possible improvement solutions were presented.Concomitantly, the proposed new BiQA2 corpus ¬ created upon the validated questions and answers from the perused samples from BiQA ¬ builds new features similar to those observed in other biomedical corpus such as the BioASQ dataset.Both BiQA and BiQA2 were applied to deep learning strategies previously submitted to the BioASQ competition to assess their performance as a source of training data. Although the results achieved with the models created using BiQA2 exhibit limited capability pertaining to the BioASQ challenge, they also show some potential to contribute positively to model training in tasks such as Document re-ranking and answering to ‘yes/no’ questions.